Installation

multigeom needs Python 3.10 to 3.13.

From PyPI

pip install multigeom

This installs the core: encoders, the Fisher metric \(J(x)\), the diffusion model, the \(G(x)\) estimator, and the information-geometric analyses. Two extras add the rest:

Install	Adds	For
pip install "multigeom[paper]"	The ON-OFF encoder (plenoptic) and the loaders of the TVSD recordings (h5py).	reproducing the paper
pip install "multigeom[all]"	Everything in paper, plus Jupyter and Modal (serverless GPUs).	the demo notebooks, everything

Check the installation:

python -c "import multigeom; print(multigeom.__version__)"

GPU

Estimating \(G(x)\) is about 15 s per image on a recent GPU and far slower on a CPU (about 15 min per image on 32 cores). Computing \(J(x)\) takes a fraction of a second and runs anywhere. Install the PyTorch build that matches your CUDA driver first, following the PyTorch instructions, then install multigeom.

From source

git clone https://github.com/multigeom/multigeom.git
cd multigeom
pip install -e ".[all]"

An editable install picks up your changes to the code without reinstalling.

Environment for the demos

The notebooks have their own repository, with a conda environment file that pins PyTorch for CUDA 12.8 and the steps to create the Jupyter kernel: see multigeom-tutorials.